Transformer diagnosis using probabilistic vibration models

Abstract

Detecting transformer failures at early basis represents enormous economical and technical advantage for a utility company. One approach reported in the literature is the vibration analysis for the detection of mechanical failures in transformers. The basic idea is the characterization of the normal vibration during the operation of the transformer, and the recognition of variations in the vibration patterns of the transformer when a failure is present. This paper presents the development of a probabilistic vibration model used for the detection of incipient failures in transformers. Vibration measurements are taken all around of the transformer tank and a probabilistic model is constructed using automatic learning algorithms developed in the Artificial Intelligence community. The models are Bayesian networks that relate probabilistically all the variables in the experiments. Later, inference algorithms are used to estimate on-line, a probability of a failure in the transformer. This project is in collaboration with Prolec General Electric, the largest constructor of transformers in North America. Experiments were carried out at Prolec GE laboratories on a power substation transformer (PST). A discussion of the experiments and their results are included in this paper. © 2009 Springer-Verlag Berlin Heidelberg.